Towards a lightweight YOLOv8n for aquaculture feeding detection: Architectural improvements for feature enhancement and computational efficiency

This study presents a novel image stitching methodology leveraging the backbone of YOLOv8 as a feature extractor, designed specifically to overcome limitations prevalent in traditional stitching techniques concerning small object detection and complex backgrounds. The proposed approach integrates a low-resolution feature map enhancement, a multi-scale feature extraction mechanism, and a deep homography estimation module. The low-resolution enhancement module utilizes an adaptive Squeeze-and-Excitation (SE) mechanism to significantly enhance the detection precision of small objects. Concurrently, the multi-scale feature extraction employs Spatial Pyramid Pooling to efficiently capture and fuse features across various resolutions, improving robustness and feature alignment in images characterized by complex textures and disparities. A core innovation of this research is the introduction of a deep learning-based homography estimation module, utilizing regression networks and Direct Linear Transformation (DLT) algorithms to compute accurate geometric transformations. Moreover, a specially designed weighted loss function emphasizes the alignment accuracy of small object regions during model training, thus reducing stitching artifacts and improving visual continuity. Experimental validation was conducted using the UDIS and MVS-Synth datasets. Results show significant improvements over baseline and comparative methods, achieving higher PSNR (26.36 and 26.42, respectively) and SSIM values (0.8414 and 0.8494, respectively), with substantial reductions in model size and computational complexity, as demonstrated by the model size (188 MB) and GFLOPs (14.5) in our experiments. Ablation studies further confirm the effectiveness of each module in enhancing stitching accuracy and computational efficiency. Future work may explore extending the model's capability for real-time application in dynamic and variable environments.

Kidney abnormalities, including cysts, tumors, and stones, affect millions worldwide and present substantial health challenges. Accurate detection and classification of these issues are essential but remain a difficult task. Computed Tomography (CT) scans, a specialized X-ray imaging technology, play a vital role in diagnosing these abnormalities. Recently, machine learning algorithms have become more prominent in these types of tasks; however, conventional machine learning techniques often struggle with data pre-processing, computational efficiency, and disease localization. To address these challenges, we have proposed an enhanced architecture of the YOLOv5 model. In this work, we primarily focused on designing a novel C3SE network method, which integrates Squeeze-and-Excitation (SE) modules with C3 (3Conv) modules within the YOLOv5, significantly improving the network's ability to detect and classify kidney disorders. Additionally, to improve adaptive optimization of feature aggregation and enhanced detection accuracy, we have employed the BiFPN. Moreover, we incorporate the SE attention mechanism to enhance multi-scale feature extraction, effectively reducing the impact of undesired input features. To validate the performance of the proposed model compared to prior approaches, we evaluated its accuracy, mean average precision (mAP), precision, F1 score, and recall, achieving 97.5%, 96.8%, 94.1%, 97%, and 96%, respectively. To ensure model interpretability and diagnostic trust, we employed Grad-CAM based explainability, visually confirming that the model accurately localizes pathological features. This promotes transparency in AI-assisted diagnosis, crucial for clinical deployment.

Brain tumors involve abnormal cell growth within or adjacent to brain tissues, necessitating precise segmentation for effective clinical decision‐making. Traditional models often face challenges in accurately delineating tumor regions, and building robust segmentation models for high‐resolution MRI data requires substantial computational power. This study presents a three‐dimensional U‐Net architecture with Squeeze and Excitation (SE) modules, called SE‐3D Brain Net, to enhance multi‐region brain tumor segmentation. The model leverages SE modules to recalibrate channel‐wise feature significance, improving segmentation accuracy across tumor subregions. Extensive experiments on datasets such as BraTS 2018 and BraTS 2020 demonstrate that the model outperforms traditional U‐Net models and various advanced methods, achieving average Dice scores of 0.86 for enhancing tumor, 0.84 for tumor core, and 0.86 for whole tumor segmentation. An ablation study further revealed the model's sensitivity to hyperparameters, identifying optimal settings for batch size, learning rate, and dropout rate. This study demonstrates the effectiveness of deep learning in accurately identifying brain tumors, emphasizing its potential to improve medical image analysis and patient outcomes significantly.

This article presents a solid color circular weft fabric defect detection method based on AYOLOv7-tiny. The aim of the development of this network is to enable real-time defect detection in the production of large circular machine fabrics. The network has a good accuracy rate, fast detection speed, and a lightweight model. In the YOLOv7-tiny network, a space-to-depth layer followed by a non-strided convolution layer is introduced to enhance the feature extraction capability, improve image sharpness, address issues such as uneven grayscale and difficult detection of minor defects, and simplify the model complexity while reducing computation. Additionally, we combined the Squeeze and Excitation (SE) and Spatial Attention Module (SAM) based on Convolutional Block Attention Module (CBAM) to construct the Hybrid Attention Module (SC), which is integrated into the YOLOv7-tiny network. The SC module increases the weight of important features, enhances the feature extraction capability, improves image segmentation, and enhances the accuracy of the network's location information. Through extensive experiments with a dataset of large circular machine solid color circular weft fabric defect collected from an industrial site, the results show that AYOLOv7-tiny has a detection accuracy (mean average precision) of 98.7%, a detection speed of up to 333 frames per second, and a computational complexity of only 4.4 GFlops, which is better than the current mainstream surface defect detection models. The detection accuracy, detection speed, and model complexity of the AYOLOv7-tiny network all meet the real-time detection requirements of the industry and have been successfully used for real-time detection of large circular machine fabric defects.

The kidney bean is an important cash crop whose growth and yield are severely affected by brown spot disease. Traditional target detection models cannot effectively screen out key features, resulting in model overfitting and weak generalization ability. In this study, a Bi-Directional Feature Pyramid Network (BiFPN) and Squeeze and Excitation (SE) module were added to a YOLOv5 model to improve the multi-scale feature fusion and key feature extraction abilities of the improved model. The results show that the BiFPN and SE modules show higher heat in the target location region and pay less attention to irrelevant environmental information in the non-target region. The detection Precision, Recall, and mean average Precision (mAP@0.5) of the improved YOLOv5 model are 94.7%, 88.2%, and 92.5%, respectively, which are 4.9% higher in Precision, 0.5% higher in Recall, and 25.6% higher in the mean average Precision compared to the original YOLOv5 model. Compared with the YOLOv5-SE, YOLOv5-BiFPN, FasterR-CNN, and EfficientDet models, detection Precision improved by 1.8%, 3.0%, 9.4%, and 9.5%, respectively. Moreover, the rate of missed and wrong detection in the improved YOLOv5 model is only 8.16%. Therefore, the YOLOv5-SE-BiFPN model can more effectively detect the brown spot area of kidney beans.

A bridge disease identification approach based on an enhanced YOLO v3 algorithm is suggested to increase the accuracy of apparent disease detection of concrete bridges under complex backgrounds. First, the YOLO v3 network structure is enhanced to better accommodate the dense distribution and large variation of disease scale characteristics, and the detection layer incorporates the squeeze and excitation (SE) networks attention mechanism module and spatial pyramid pooling module to strengthen the semantic feature extraction ability. Secondly, CIoU with better localization ability is selected as the loss function for training. Finally, the K-means algorithm is used for anchor frame clustering on the bridge surface disease defects dataset. 1363 datasets containing exposed reinforcement, spalling, and water erosion damage of bridges are produced, and network training is done after manual labelling and data improvement in order to test the efficacy of the algorithm described in this paper. According to the trial results, the YOLO v3 model has enhanced more than the original model in terms of precision rate, recall rate, Average Precision (AP), and other indicators. Its overall mean Average Precision (mAP) value has also grown by 5.5%. With the RTX2080Ti graphics card, the detection frame rate increases to 84 Frames Per Second, enabling more precise and real-time bridge illness detection.

MSTSENet: Multiscale Spectral–Spatial Transformer with Squeeze and Excitation network for hyperspectral image classification

High-speed laser cladding is a critical surface modification technique, where the melt pool state critically determines the quality of the cladding layer. To address the challenges of inaccurate monitoring and the lack of real-time feedback control during the cladding process, this study proposes an improved YOLOv5-based object detection model, named YOLO-MPID. The model incorporates the Squeeze-and-Excitation (SE) module and the Convolutional Block Attention Module (CBAM) to enhance the extraction of complex edge structures and semantic features in melt pool images. Ablation studies were conducted to evaluate the individual and combined effects of SE and CBAM on detection performance. Comparative experiments under varying laser power conditions demonstrated that YOLO-MPID achieves superior robustness and accuracy compared to the baseline YOLOv5 and other mainstream detection algorithms. Experimental results show that the proposed model achieves a mean Average Precision (mAP at 0.5) of 97.42% and a real-time inference speed of 121.71 FPS. Visual analysis further supports the quantitative findings. In summary, YOLO-MPID provides an effective solution for real-time melt pool state detection and quality control in practical industrial scenarios, offering robust technical support for advancements in image science and optical process control.

The current challenges in Synthetic Aperture Radar (SAR) ship detection tasks revolve around handling significant variations in target sizes and managing high computational expenses, which hinder practical deployment on satellite or mobile airborne platforms. In response to these challenges, this research presents YOLOv7-LDS, a lightweight yet highly accurate SAR ship detection model built upon the YOLOv7 framework. In the core of YOLOv7-LDS's architecture, we introduce a streamlined feature extraction network that strikes a delicate balance between detection precision and computational efficiency. This network is founded on Shufflenetv2 and incorporates Squeeze-and-Excitation (SE) attention mechanisms as its key elements. Additionally, in the Neck section, we introduce the Weighted Efficient Aggregation Network (DCW-ELAN), a fundamental feature extraction module that leverages Coordinate Attention (CA) and Depthwise Convolution (DWConv). This module efficiently aggregates features while preserving the ability to identify small-scale variations, ensuring top-quality feature extraction. Furthermore, we introduce a lightweight Spatial Pyramid Dilated Convolution Cross-Stage Partial Channel (LSPHDCCSPC) module. LSPHDCCSPC is a condensed version of the Spatial Pyramid Pooling Cross-Stage Partial Channel (SPPCSPC) module, incorporating Dilated Convolution (DConv) as a central component for extracting multi-scale information. The experimental results show that YOLOv7-LDS achieves a remarkable Mean Average Precision (mAP) of 99.1% and 95.8% on the SAR Ship Detection Dataset (SSDD) and the NWPU VHR-10 dataset with a parameter count (Params) of 3.4 million, a Giga Floating Point Operations Per Second (GFLOPs) of 6.1 and an Inference Time (IT) of 4.8 milliseconds. YOLOv7-LDS effectively strikes a fine balance between computational cost and detection performance, surpassing many of the current state-of-the-art object detection models. As a result, it offers a more resilient solution for maritime ship monitoring.

An improved YOLOv5 model based on visual attention mechanism: Application to recognition of tomato virus disease

Tea yield estimation using UAV images and deep learning

Discrimination between normal (fresh/non-frozen) and frozen-thawed beef is crucial for ensuring food safety. This paper proposed a novel, non-destructive and real-time you only look once for normal and frozen-thawed beef discrimination (YOLO-NF) model using deep learning techniques. The simple, parameter-free attention module (SimAM) and the squeeze and excitation (SE) attention mechanism were introduced to enhance the model’s performance. A total of 1200 beef samples were used, with their images captured by a charge-coupled device (CCD) camera. In the model development, specifically, the training set comprised 3888 images after data augmentation, while the validation set and test set each included 216 original images. Experimental results on the test set showed that the YOLO-NF model achieved precision, recall, F1-Score and mean average precision (mAP) of 95.5%, 95.2%, 95.3% and 98.6%, respectively, significantly outperforming YOLOv7, YOLOv5 and YOLOv8 models. Additionally, gradient-weighted class activation mapping (Grad-CAM) was adopted to interpret the model’s decision basis. Moreover, the model was deployed on the web interface for user convenience, and the discrimination time on the local server was 0.94 s per image, satisfying the requirements for real-time processing. This study provides a promising technique for high-performance and rapid meat quality assessment in food safety monitoring systems.

This study introduces an improved YOLOv8 model YOLO-FSL for rapid detection of plum maturity under natural conditions. Firstly, the Faster Net Block module was introduced into the C2f module of the YOLOv8s backbone network, forming the C2f_Faster module; Secondly, squeeze and excitation (SE) attention mechanism after each C2f_Faster module was added in the backbone network; Finally, the improved YOLOv8s model was pruned using channel pruning algorithm to simplify the detection model and ensure detection efficiency. The experimental results show that the mean average precision (mAP) and model size of the improved YOLO-FSL model is 91.2% and 4.1 MB (Mega Byte), respectively, which reduces the number of parameters by 82.57% and improves the mAP by 1.6% compared with the original network YOLOv8s. Compared with the mainstream target detection algorithms of YOLOv3-tiny, YOLOv5s, YOLOv6s, and YOLOv7-Tiny, the mAP of the YOLO-FSL model is improved by 4.4%, 2.0%, 1.0%, and 1.6%, respectively. The YOLO-FSL model has a higher detection accuracy, and at the same time reduces the amount of computation and the number of parameters, and lowers the hardware dependence, which provides a technical reference for the automated picking of plums.

Background: The generation of synthetic CT images from MRI scans represents a crucial step toward enabling MRI-only clinical workflows and supporting multi-modal integration in medical imaging, particularly in radiotherapy planning. Despite significant advancements in deep learning models, many current methods still struggle to reconstruct high-density structures, especially bone, and exhibit limited accuracy in density values. This shortcoming is largely attributed to the passage of excessive or noisy features through skip connections in the traditional U-Net architecture, which degrade the quality of information transmitted to the decoder, negatively impacting the clarity of anatomical boundaries and the pixel-wise accuracy of the resulting synthetic image. Methods: In this work, we propose an enhanced 3D U-Net architecture in which the Convolutional Block Attention Module (CBAM) is systematically integrated within each skip connection. The CBAM sequentially applies channel and spatial attention to adaptively reweight encoder feature maps before fusion with the decoder, thereby emphasizing anatomically relevant structures while suppressing irrelevant feature propagation. The model was trained and evaluated on the SynthRAD2023 (Task 1-Brain) MRI-CT dataset. To rigorously assess the contribution of the attention mechanism, a dedicated ablation study was conducted comparing three variants: 3D U-Net with Squeeze-and-Excitation (SE), Coordinate Attention (CA), and the proposed CBAM module. Performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Cross-Correlation (NCC). Results: The ablation study demonstrated that the CBAM-enhanced model consistently outperformed both SE- and CA-based variants across all quantitative metrics. Specifically, the proposed method achieved an MAE of 38.2±5.4 HU and an RMSE of 51.0±12.0 HU, representing the lowest reconstruction errors among the evaluated models. In addition, it obtained a PSNR of 29.45±2.10 dB, SSIM of 0.940±0.031, and NCC of 0.967±0.015, indicating superior structural preservation and strong voxel-wise correspondence between synthesized and reference CT volumes. These results confirm that the sequential integration of channel and spatial attention provides a statistically and practically meaningful improvement for high-fidelity MRI-to-CT synthesis. Conclusions: Generating high-resolution brain CT images from brain MRI scans using a 3D U-Net network enhanced with a CBAM module can contribute to supporting the clinical workflow by providing additional diagnostic data without the need for extra radiological examinations, thereby enhancing diagnostic efficiency and reducing radiation exposure. This technique helps reduce patient exposure to radiation and improves accessibility in resource-limited settings. Furthermore, this method is valuable for retrospective studies, surgical planning, and image-guided therapy, where complete multi-modal data may not always be available.

The heart is a relatively complex non-rigid motion organ in the human body. Quantitative motion analysis of the heart takes on a critical significance to help doctors with accurate diagnosis and treatment. Moreover, cardiovascular magnetic resonance imaging (CMRI) can be used to perform a more detailed quantitative analysis evaluation for cardiac diagnosis. Deformable image registration (DIR) has become a vital task in biomedical image analysis since tissue structures have variability in medical images. Recently, the model based on masked autoencoder (MAE) has recently been shown to be effective in computer vision tasks. Vision Transformer has the context aggregation ability to restore the semantic information in the original image regions by using a low proportion of visible image patches to predict the masked image patches. A novel Transformer-ConvNet architecture is proposed in this study based on MAE for medical image registration. The core of the Transformer is designed as a masked autoencoder (MAE) and a lightweight decoder structure, and feature extraction before the downstream registration task is transformed into the self-supervised learning task. This study also rethinks the calculation method of the multi-head self-attention mechanism in the Transformer encoder. We improve the query-key-value-based dot product attention by introducing both depthwise separable convolution (DWSC) and squeeze and excitation (SE) modules into the self-attention module to reduce the amount of parameter computation to highlight image details and maintain high spatial resolution image features. In addition, concurrent spatial and channel squeeze and excitation (scSE) module is embedded into the CNN structure, which also proves to be effective for extracting robust feature representations. The proposed method, called MAE-TransRNet, has better generalization. The proposed model is evaluated on the cardiac short-axis public dataset (with images and labels) at the 2017 Automated Cardiac Diagnosis Challenge (ACDC). The relevant qualitative and quantitative results (e.g., dice performance and Hausdorff distance) suggest that the proposed model can achieve superior results over those achieved by the state-of-the-art methods, thus proving that MAE and improved self-attention are more effective and promising for medical image registration tasks. Codes and models are available at https://github.com/XinXiao101/MAE-TransRNet.